Charley Sanchez
Photo of Charley Sanchez

Charley Sanchez

PhD student, Computer Science
University of Maryland, Baltimore County

I'm a PhD student at UMBC, where I work with Prof. Dong Li in the Future Sensing and Interaction (FSI) Lab. I study the security of foundation models for physiological signals such as EEG, ECG, and wearable data. My current work examines how multimodal health models respond to realistic sensor faults and attacks, and how their design choices shape the damage.

Before UMBC, I worked on mental health modeling from wearable data with Prof. Anind K. Dey at Georgia Tech, causal machine learning with Prof. Wenbo Wu at Johns Hopkins, and privacy-preserving perception with Prof. Nader Sehatbakhsh at UCLA, where I did my M.Eng. I have a B.S. in Physics from UCSB, and after graduating I worked in a clinical diagnostic lab, which is a big part of why I ended up in healthcare ML.

charleysanchez@umbc.edu CV GitHub LinkedIn

News

Research

Robustness of multimodal physiological foundation models

UMBC, with Prof. Dong Li · 2026 to present

Foundation models trained on large collections of biosignals, such as EEG, ECG, EOG, and EMG, are increasingly proposed for clinical and wearable health monitoring. These systems combine several sensors, so an attacker, or simply a faulty electrode, can target a single input stream rather than the whole model. My research builds a threat-model-driven evaluation of these systems: realistic attacks defined at the level of the physical signal, applied consistently across models with different architectures, and measured with statistically careful, subject-level evaluation. The goal is to understand which design decisions make multimodal health models fragile or resilient, and what that means for deploying them safely.

Adversarial machine learning · Health AI security · Multimodal fusion · Biosignals · Sleep staging

AI for mental health

Georgia Tech, with Prof. Anind K. Dey · 2026

I built modeling pipelines for mental health assessment from wearable and behavioral time-series data. I investigated encoding the time series as images (Markov Transition Fields and Gramian Angular Fields) so convolutional and multimodal models could learn from them, and designed a gradient-boosted feature selection step to keep CNN training focused when data was limited. I finished the project and handed off the codebase and documentation to the next research cohort.

Causal machine learning

Johns Hopkins, with Prof. Wenbo Wu · 2025–2026

Double machine learning estimates treatment effects by first fitting flexible “nuisance” models, and I studied how the choice of deep architecture for those models (ResMLP, DCN, Transformers) affects the bias, variance, and stability of multi-treatment effect estimates. Across 500–2,000+ configurations, residual MLP blocks introduced the least covariate distortion, and a better-designed gamma model recovered true treatment effects 10–22% more accurately in nonlinear synthetic settings. I also started a literature review on cost-supervised embeddings of diagnosis codes (ICD/CPT) for estimating patient costs in Medicare Advantage plans.

Privacy-preserving perception

UCLA, with Prof. Nader Sehatbakhsh · 2025

In the Secure Systems and Architectures Lab, I worked on real-time face anonymization for camera-equipped robots, with Nokia Bell Labs as an industry partner. On a Jetson Orin Nano, I designed mosaic and noise-based anonymization that cut compute by over 92% while keeping the visual signal navigation depends on, and built the evaluation pipeline against Segment Anything masks. This work led to the Argus manuscript below.

Papers and reports

  1. Argus: Real-Time Privacy-Preserving Video Streaming for Delivery Robots H. Khalili, P. Do, A. Dabiran, C. Sanchez, A. Hafemeister, V. Nguyen, K. Apicharttrisorn, N. Sehatbakhsh Manuscript in preparation
  2. PAVAC: Privacy-Aware Vehicular Autonomous Computation P. Do, A. Dabiran, V. Nguyen, A. Hafemeister, C. Sanchez Technical report, UCLA, 2025 [pdf] [code] [video]
  3. Bridging the Generalization Gap in sEMG Keystroke Recognition with LSTM-Based Architectures A. Jain, C. Sanchez Technical report, UCLA, 2025 [pdf] [code]
  4. Spurious Correlation Detection in Natural Language Processing M. Berker, N. Huey, C. Sanchez Technical report, UCLA, 2024 [pdf] [code]

Projects

Experience

Education